US20160188725A1 - Method and System for Enhanced Content Recommendation - Google Patents

Abstract

Method, system, and programs for providing content recommendation are disclosed. A first set of candidate content items may be generated based on a user profile, and a second set of candidate items may be generated based on the likelihood that the user will click a corresponding candidate content item in the second set. The candidate content items in the first and second sets may be ranked together using a learning model and presented to the user as content recommendations based on their rankings. The likelihood that the user will click a given candidate content item in the second set may be estimated based on similarities between the given content item and content items related to the given content item. Such a similarity may be computed based on activities performed by users who have viewed both the given content item and a related content item.

Description

BACKGROUND
• 1. Technical Field
• The present teaching relates to providing content. Specifically, the present teaching relates to methods and systems for providing online content.
• 2. Discussion of Technical Background
• The Internet has made it possible for a user to electronically access virtually any content at any time and from any location. With the explosion of information, it has become more and more important to provide users with information that is relevant to the user and not just information in general. Further, as the Internet has become an important source of information for millions of users, including entertainment, and/or social connections, e.g., news, social interaction, movies, music, etc., it is critical to provide users with information they find valuable.
• Efforts have been made to allow users to readily access relevant and on the point content. For example, topical portals have been developed that are more subject matter oriented as compared to generic content gathering systems such as traditional search engines. Example topical portals include portals on finance, sports, news, weather, shopping, music, art, film, etc. Such topical portals allow users to access information related to subject matters that these portals are directed to. Users have to go to different portals to access content of certain subject matter, which is not convenient and not user centric.
• Another line of efforts to enable users to easily access relevant content is via personalization, which aims at understanding each user's individual likings/interests/preferences so that an individualized user profile for each user can be set up and can be used to select content that matches a user's interests. The underlying goal is to meet the minds of users in terms of content consumption. User profiles traditionally are constructed based on users' declared interests and/or inferred from, e.g., users' demographics. There have also been systems that identify users' interests based on observations made on users' interactions with content. A typical example of such user interaction with content is click through rate (CTR).
• CTR may have been the most commonly used measure to estimate users' interests. However, CTR is not the only type of information adequate to capture information reflecting users' interests particularly given that other different types of activities that a user may perform may also indicate or implicate user's interests. In addition, user reactions to content usually represent users' short term interests. Such observed short term interests, when acquired in piece meal, as traditional approaches often do, can only lead to reactive, rather than proactive, services to users. Although short term interests are important, they may not be sufficient to reach an understanding of the more persistent long term interests of a user, which are crucial in terms of user retention. Most user interactions with content represent short term interests of the user so that relying on such short term interest behavior makes it difficult to expand the understanding of the increasing range of interests of the user. When this is in combination with the fact that such collected data is always the past behavior and collected passively, it creates a personalization bubble, making it difficult, if not impossible, to discover other interests of a user unless the user initiates some action to reveal new interests.
• Yet another line of effort to allow users to access relevant content is to pooling content that may be of interest to users in accordance with their interests. Given the explosion of information on the Internet, it is not likely, even if possible, to evaluate all content accessible via the Internet whenever there is a need to select content relevant to a particular user. Thus, realistically, it is needed to identify a subset or a pool of the Internet content for individual users or a subgroup of users who share interests based on some criteria so that content can be selected from this pool and recommended to users based on their interests for consumption.
• Conventional approaches to creating such a subset of content are application centric. Each application carves out its own subset of content in a manner that is specific to the application. For example, Amazon.com may have a content database related to products and information associated thereof created/updated based on information related to its own users and/or interests of such users exhibited when they interact with Amazon.com. Based on knowledge about individual users, Amazon.com may generate a pool for each user based on their purchasing preferences. Facebook may also have its own subset of content, generated in a manner not only specific to Facebook but also based on user interests exhibited while they are active on Facebook. As a user may be active in different applications (e.g., Amazon.com and Facebook) and with each application, they likely exhibit only part of their overall interests in connection with the nature of the application. Given that, each application can usually gain understanding, at best, of partial interests of users, making it difficult to develop a subset of content that can be used to serve a broader range of users' interests.
• Yet another line of effort is directed to personalized content recommendation, i.e., selecting content from a content database based on the user's personalized profiles and recommending such identified content to the user. Conventional solutions focus on relevance, i.e., the relevance between the content and the user's past interests. For example, a user's profile indicating a set of features (e.g., terms, phrases, topics, categories) of content viewed by the user in the past is typically extracted for comparison with the features in the contents. However, such relevance based content recommendation techniques are limited in that they require feature extraction from both the user profile and contents to be accurate so that features in the contents may be matched to the features in the user profile for recommendation.
• Another line of effort is directed to recommending contents based on user activity information. For example, some conventional systems analyze user activities and generate a bipartite graph indicating whether users viewed certain content items. For instance, an edge may be established in the bipartite graph between user A and content item A to indicate user A viewed content item A, another edge may be established in the bipartite graph between user A and content item C to indicate user A viewed content item C, a third edge may be established in the bipartite graph between user B and content item C to indicate user B viewed content item C, and so on. Based on this bipartite graph, these systems may recommend content item C to a given user who viewed content item A because user A, who is deemed to be similar to the given user by those systems, viewed content items A and C. To these systems, user A's viewing of content items A and C correlates the given user's interest and thus content item C should be recommended to the user after the given user viewed content item A. However, such an approach is limited because the bipartite graph may not accurately capture user A's interest. For example, lacking of an edge between user A and a content item, say content item D, does not necessarily mean user A is not interested in content item D. There could be a number of other reasons why user A did not view content item D—for example, content item D may not have been even presented to user A or user A may not have discovered content item D. As another example, the edge between user A and content item C indicating user A viewed content item C may also not necessarily mean user A is interested in content item C. For instance, user A may have viewed content item C merely for a quick overview and decides content item C is not of interest to him/her-self. On the other hand, other user activities such as user click activities during viewing of the content items may provide further insights into user interests in content items. They may be used to enhance and enrich the conventional user-activity based content recommendation approach.
• Accordingly, there is at least a need to enhance conventional content recommendation techniques.
SUMMARY
• The present teaching relates to providing content. Specifically, the present teaching relates to methods and systems for providing online content.
• In one example, a method, implemented on a machine having at least one processor, storage, and a communication platform connected to a network, for presenting providing content recommendations.
• In accordance with the present teaching, for recommending content items to a user, a first set of candidate content items may be generated based on a user profile, and a second set of candidate items may be generated based on the likelihood that the user will click a corresponding candidate content item in the second set. The candidate content items in the first and second sets may be ranked together using a learning model and presented to the user as content recommendations based on their rankings.
• For generating the first set of candidate content items for recommendation to the user, information in the user's profile indicating content features that have been viewed by the user may be obtained. The number of times the user has viewed a particular one of these content features may be compared with average number of times other users have viewed that content feature. Such a comparison may be carried out for each content feature that has been viewed by the user. The content features that have been viewed by the user may then be ranked based on the results of such comparisons. A number of content features that have been viewed by the user may be selected based on their rankings to reflect the user's interest in content features. The first set of the candidate content items may be generated from a content storage based on the number of selected content features.
• For generating the second set of candidate content items, the likelihood that the user will click a given candidate content item may be estimated based on similarities between the given content item and content items related to the given content item and viewed by the user previously. A similarity between the given content item and a related content item may be generated based on activities performed by users who have viewed both the given content item and the related content item. The user activities may include clicking, typing, scrolling, dwelling, forwarding, commenting, and/or any other types(s) of activities by those users during viewing of the given content item and the related content item. In implementations, for computing such a similarity, a user activity vector for the given candidate content item may be generated. The values of the generated user activity vector may indicate weighted user activities performed by corresponding users during viewing of the given candidate content item. A user activity vector for the related content item may be similarly generated. The similarity between the given candidate content item and the related content item may be estimated by comparing the two user activity vectors. Similarities between the given candidate content item and each related content item may be estimated in this fashion and aggregated. Based on the aggregated similarities between the given candidate content item and the related content items and whether the user has clicked the related content items, the likelihood that the user will click the given candidate content item may be determined. If the likelihood is high enough, the given candidate content item may be included in the second set of candidate content items.
• The candidate content items in the first and second sets may be ranked together using a learning model. In some implementations, the learning model may be trained using user information, content information, user-content cross information, and/or any other type(s) of information. Candidate content items in the first and second sets may be presented to the user as content recommendations based on their rankings determined using the learning model.
• Other concepts relate to software for implementing the enhanced content recommendations. A software product, in accord with this concept, includes at least one machine-readable non-transitory medium and information carried by the medium. The information carried by the medium may be executable program code data regarding parameters in association with a request or operational parameters, such as information related to a user, a request, or a social group, etc.
• In one example, a machine readable and non-transitory medium having information recorded thereon for recommending content items, where when the information is read by the machine, causes the machine to obtain a user profile characterizing interests of the user; generate a first set of candidate content items based on the user profile; generate a second set of candidate content items based on a likelihood that the user clicks a corresponding candidate content item in the second set, wherein each likelihood is estimated based on similarities between the candidate content items in the second set and one or more content items that were previously viewed by the user; rank each of the candidate content items in the first set and the second set; and provide, based on the rankings, the candidate content items in the first and second sets as content recommendations to the user.
• Additional advantages and novel features will be set forth in part in the description which follows, and in part will become apparent to those skilled in the art upon examination of the following and the accompanying drawings or may be learned by production or operation of the examples. The advantages of the present teachings may be realized and attained by practice or use of various aspects of the methodologies, instrumentalities and combinations set forth in the detailed examples discussed below.
BRIEF DESCRIPTION OF THE DRAWINGS
• The methods, systems and/or programming described herein are further described in terms of exemplary embodiments. These exemplary embodiments are described in detail with reference to the drawings. These embodiments are non-limiting exemplary embodiments, in which like reference numerals represent similar structures throughout the several views of the drawings, and wherein:
• FIG. 1 depicts an exemplary system for personalized content recommendation, according to an embodiment of the present teaching;
• FIG. 2 is a flowchart of an exemplary process for providing content recommendation, according to various embodiments of the present teaching;
• FIG. 3 illustrates one example of the user-profile based candidate item selection unit shown inFIG. 1, according to an embodiment of the present teaching;
• FIG. 4 illustrates an exemplary method that may be implemented by the relevance-based content similarity determination module shown inFIG. 3, according to various embodiments of the present teaching;
• FIG. 5 illustrates one example of a user-activity based candidate item selection unit shown inFIG. 1, according to an embodiment of the present teaching;
• FIG. 6 conceptually illustrates one example of user activity information associated with a set content items;
• FIG. 7 conceptually illustrates how a similarity between two content items can be determined based on their user activity information, according to an embodiment of the present teaching;
• FIG. 8 illustrates another example how a similarity between two content items can be determined based on their user activity information, according to an embodiment of the present teaching;
• FIG. 9 conceptually illustrates how a similarity between content items may be determined based on user activity vectors associated with the content items, according to an embodiment of the present teaching;
• FIG. 10 is a flowchart of an exemplary process for determining a similarity between two content items, according to one embodiment of the present teaching.
• FIG. 11 is a flowchart of another exemplary process for determining a similarity between two content items, according to one embodiment of the present teaching;
• FIG. 12 is a flowchart of an exemplary process for estimating likelihood that a user will click a given content item, according to an embodiment of the present teaching;
• FIG. 13 is an exemplary system diagram of a machine learning engine shown inFIG. 1, according to an embodiment of the present teaching;
• FIG. 14 illustrates examples of user side features that may be used to train a learning model;
• FIG. 15 illustrates examples of content side features that may be used to train a learning model;
• FIG. 16 illustrates an exemplary system diagram of the unified ranking unit shown inFIG. 1, according to one embodiment of the present teaching.
• FIG. 17 depicts the architecture of a mobile device which can be used to implement a specialized system incorporating the present teaching; and
• FIG. 18 depicts the architecture of a computer which can be used to implement a specialized system incorporating the present teaching.
DETAILED DESCRIPTION
• In the following detailed description, numerous specific details are set forth by way of examples in order to provide a thorough understanding of the relevant teachings. However, it should be apparent to those skilled in the art that the present teachings may be practiced without such details. In other instances, well known methods, procedures, components, and/or circuitry have been described at a relatively high-level, without detail, in order to avoid unnecessarily obscuring aspects of the present teachings.
• The present teaching relates to recommending on-line content to a user. Particularly, the present teaching relates to a system, method, and/or programs for personalized content recommendation that addresses the shortcomings associated the conventional content recommendation solutions in personalization, content database, and recommending personalized content. The present teaching may be realized as a specialized and networked system by utilizing one or more computing devices (e.g., mobile phone, personal computer, etc.) and network communications (wired or wireless).
• In accordance with the present teaching, a first set of candidate content items may be generated based on a user profile. The user profile may comprise information indicating content features that have been viewed by the user. The individual content features in the user profile may be ranked by comparing a number of times the user viewed a corresponding content feature and an average number of times this content feature was viewed by a general population of users. A number of content features may be selected based on the rankings of the content features. Based on the selected content features, the first set of candidate content items may be generated for recommendation to the user.
• In accordance with the present teaching, a second set of candidate items may be generated based on the likelihood that the user will click a corresponding candidate content item in the second set. In some implementations, the likelihood that the user will click a given candidate content item may be estimated based on similarities between the given candidate content item and related content items that were viewed by the user previously. A similarity between the given candidate content item and a related content item may be generated based on activities by users who have viewed both the given candidate content item and the related content item. The activities may include clicking, typing, scrolling, dwelling, forwarding, commenting, and/or any other types(s) of activities by those users during viewing of the given candidate content item and the given related content item. In implementations, for computing such a similarity, a user activity vector for the given candidate content item may be generated having various weights associated with corresponding activities that were performed by corresponding users during viewing of the given candidate content item. A user activity vector for a given related content item may be similarly generated. The similarity between the given candidate content item and a related content item may then be computed by comparing the two user activity vectors. The similarities between the given candidate content item and related content items may be aggregated. The likelihood that the user will click the given candidate content item may be determined based on the aggregated similarities and whether the related content items were clicked by the user. If the likelihood is high enough, the given candidate content item may be included in the second set of candidate content items.
• After the first and second sets of candidate content items are generated, in accordance with the present teaching, the candidate content items in the first and second sets may be ranked. In some examples, a learning model may be used to rank the candidate content items in the first and second sets. The learning model may be trained using a set of training data including user features, content features, user-content cross features, and/or any other features. These features may be gathered and fed into the learning model to train the learning model. The trained learning model may then be used to rank the candidate content items in the first and second sets for presentation to the user.
• FIG. 1 depicts anexemplary system10 for personalized content recommendation, according to an embodiment of the present teaching. As shown in this example,system10 may comprise a personalizedcontent recommendation module100, amachine learning engine108, storage112, and/or any other components. As shown,system10 may be operatively connected to or include content source(s)120, knowledge archives130, third party platform(s)140, andexternal resources150. Content source(s)120 may include any source of on-line content such as on-line news, published papers, blogs, on-line tabloids, magazines, audio content, image content, and video content. Examples of contents provided by content source(s)120 may include contents from content provider such as Yahoo! Finance, Yahoo! Sports, CNN, and ESPN; may include textual, multi-media contents and/or any other form of content suitable for user interaction or viewing via the world wide web; may include social media contents accessible on a social media site, such as Facebook, twitter, Reddit, and/or or any other social media site; may include licensed contents from providers such as AP and Reuters; may include contents crawled and indexed from various sources on the Internet; and/or may include any other types of contents.Content sources120 provide a variety of content tosystem10 for providing personalized content recommendation. It is understood thatcontent sources120 may include sources of contents that are not necessarily accessible via Internet. For example,content sources120 may include sources of contents accessible via a local communication network, such as an intranet.
• Knowledge archives130 may include any depository of knowledge related information, including an on-line encyclopedia such as Wikipedia or indexing system such as an on-line dictionary. Knowledge archives130 may be utilized for its content as well as its categorization or indexing systems. Knowledge archives130 may provide classification system to assist with the classification of both the preferences of user as well as classification of content. Knowledge archives130, such as Wikipedia may have hundreds of thousands to millions of classifications and sub-classifications, which may be organized in hierarchies. Those classifications may be used bysystem10 to determine or select personalized content to be recommended to a user. For example, they may helpsystem10 to understand how one category of contents relates to another category of contents. They may helpsystem10 to maneuver between higher levels on the hierarchy without having to move up and down the subcategories. The categories or classification structure inKnowledge archives130 may be used for constructing multidimensional content vectors and multidimensional user profile vectors, which may be utilized by personalizedcontent recommendation module100 to match personalized content.
• The third party platform(s)140 may include any platform provided by a third party service provider affiliated with, e.g., thesystem10. Examples of third party platforms may include, but not limited to, social networking sites like Facebook, Twitter, LinkedIn, Google+, mail servers such as Gmail, online community such as America Online, and/or any other type(s) of third party platform(s).Third party platforms140 may providesystem10 contents as well as information that may be utilized to understand insights of user preferences and behaviors. Such insights may be used to helpsystem10 to recommend personalized content to individual users. As illustration, without limitation, information from a third party platform on user activities performed on the third party platform may reveal users' interests exhibited on the third party platform, content that the users consumed on the third party platform, and/or any other user information from the third party platform that may be used to enhance the personalization of content stream provided to a particular user ofsystem10. For example, when information about a large user population can be accessed from one or morethird party platforms140,system10 can rely on data about a large population to establish a baseline interest profile to estimate the interests of individual users more precise and reliable, e.g., by comparing interest data with respect to a particular user with the baseline interest profile which will capture the user's interests with a high level of certainty.
• External resources150 may provide various types of content tosystem10, including but not limited to, streaming contents, static contents, and sponsored contents.External resources150 may include any party associated with an advertising entity that is associated withsystem10. Such an external resource may provide content to be included in the recommendation to a user, including, but not limited to, advertisements from an advertisement content database (not shown), information related to target audience for each advertisement, and/or information related to advertisement taxonomy that may be used to identify advertisement in appropriate categories. Advertising contents may be presented either as a part of a content stream or a standalone advertisement such as a pop-up window, and can be placed either according to a designated ad-space or strategically around or within the associated content stream. In some implementations, appropriate advertising content may be selected by the personalizedcontent recommendation module100 in order for it to be inserted into a content stream. The personalization of advertising content may be based on a variety types of information such as user profile110 or user activities. In some implementations, content taxonomy, advertisement taxonomy, and user information, including user profile or activities, may all be used to determine appropriate advertisement given the content related to certain subject matter to be presented to a specific user that has certain preferences.
• In some implementations,system10 may be configured to identify and/or process contents from content source(s)120, knowledge archives130, 3^rdparty platforms140, and/orexternal resources150 for identifying personalized content recommendations to user. For instance, when new content is identified fromcontent source120 or third party platform(s)140,system10 may be configured to analyze the new content to understand concepts or subject matter embodied in the new contents. Such obtained concepts can be mapped to one or more categories in a content taxonomy (not shown) associated withsystem10. The content taxonomy may include an organized structure of concepts or categories of concepts and it may contain a few hundred classifications of a few thousand. In addition, the knowledge archives130 may also provide millions of concepts, which may or may not be structured in a similar manner as the content taxonomy. The content taxonomy and, in some examples, the knowledge archives130 may be used as a universal interest space. Concepts estimated from the new content can be mapped to this universal interest space and a high dimensional content feature vector may be constructed for the new content and used to characterize the new content. Similarly, for a user, a personal interest profile may be constructed, mapping the user's interests, characterized as concepts, to the universal interest space so that a high dimensional user feature vector can be constructed with the user's interests levels populated in the general vector.
• The personalizedcontent recommendation module100 may be configured to provide personalized content to a user. In some implementations, operation(s) of personalizedcontent recommendation module100 may be triggered when the user engagessystem10 through aclient device160 associated with user.Client device160 may include a smart phone, a tablet, a laptop computer, a desktop computer, a wearable device, an embedded computing device (such as in a car), a game console, a set-top box, a printer, or any other type of client device. User information includes, e.g., user id, login information, session information, user activity information (e.g., click, typing, browsing, and/or any other type user activities), or search information (e.g., search terms entered by the user to system10). Some user information reflects user's interests. Information related to any other user may also be received from theclient device160 operated by such users so that interests of a group of users may also be collected and utilized. User may interact withsystem10 via any suitable communications channel including a wired link and/or a wireless link. The user information received fromuser devices160 may be utilized by the personalizedcontent recommendation module100 to locate content that is customized with respect to individual users for recommending to the users on theirrespective client device160. In various implementations, the personalized content recommended to the user may include personalized content stream presented in an application running on theclient device160.
• In some implementations, as shown in this example, the personalizedcontent recommendation module100 may include aunified ranking unit102, a user activity-based candidatecontent selection unit104, a user-profile based candidate content selection unit106, amachine learning engine108, and/or any other appropriate components. As shown, the user-profile based candidate content selection unit106 may be configured to identify content appropriate for the user based on the profile related to the user. In operation, the user-profile based candidate content selection unit106 receives a user profile110 of a user (user #1 in this example) and various content items from acontent database114 in the storage112 as inputs. The user-profile based candidate content selection unit106 may be configured to generate a first set of candidate content items based on these inputs.
• The user-activity based candidatecontent selection unit104 may be configured to identify appropriate content for a user based on activities of the user. In operation, the user-activity based candidatecontent selection unit104 may receive information related to user activities, e.g., from theuser activity database116 and select a second set of candidate content items from thecontent database114 that are deemed as appropriate given the observed user activities. Theunified ranking unit102 may be configured to receive the first and second candidate content items (from the user-profile based candidate content selection unit106 and user-activity based candidatecontent selection unit104, respectively), and rank the content items received based on a learning model provided by themachine learning engine108. As will be described later, themachine learning engine108 may be configured to train learning model(s), which may be used by theunified ranking unit102 to rank the candidate content items in the first and second sets content items to be recommended to a user.
• Storage112 may be configured to store various types of information that may be used to assist personalizedcontent recommendation module100 to provide personalized content recommendations. As shown in this example, storage112 may include, but is not limited to, acontent database114 configured to store information related to content items, and auser activity database116 configured to store information related to user activities.Content database114 may include content items that can be presented to the users and/or Meta information associated therewith. For example, content items may include, but not limited to, articles, movies, music clips, photos, animation, sounds, games, interactive applications (e.g., java applets), contents in an app on a portable device, virtual reality contents, and/or any other type of content items. In some implementations, thecontent database114 may be structured so that it may store personalized content pool for individual users. In this case, content items in the content pool may be generated and retained with respect to individual users. In implementations, the content pools incontent database114 may be organized as a tiered system with both the general content pool and personalized individual content pools for different users. In some implementation, in a content pool for a user, the content items itself may not be physically present but is operational via links, pointers, or indices which provide references to where the actual content is stored in the general content pool.
• Content database114 may be dynamically updated bysystem10. Content item in the content database may come and go and decisions may be made based on the dynamic information of the users, the content itself, as well as other types of information. For example, without limitation, when the performance of content deteriorates, e.g., low level of interests exhibited from users,system10 may decide to purge it from thecontent database114. When content becomes stale or outdated, it may also be removed from thecontent database114. When there is a newly detected interest from a user,system10 may fetch new content aligning with the newly discovered interests and then update thecontent database114 by, e.g., insert the content into a relevant content pool in thedatabase114.
• Theuser activity database116 may be configured to store information related to activities that user conducted. User events may be an important source of observations as to content performance and user interest dynamics. Information of user activities may be analyzed bysystem10 in determining inferred user interests or dynamics of the user interests. When new interests are identified for a user, content appropriate for the new interests may be identified and used to updatecontent database114. When fetching new content (e.g., fromcontent sources120, knowledge archives130, 3^rdparty platform(s)140, and/or external resources150),system10 may invoke a content crawler (not shown in this example) to gather new content, which may then be analyzed and evaluated as to its quality and performance beforesystem10 decides whether it will be included in the content pool associated with the user in thecontent database114. Content may also be removed fromcontent database114 when it is no longer relevant, e.g., either because other users are not considering it to be of high quality or because it is no longer timely. As content is constantly changing, updatingcontent database114 may also be performed constantly or dynamically to change or update thecontent database114 so that potentially high relevant, high quality, and timely content can be made available to the personalizedcontent recommendation module100 for recommendation.
• Theuser activity database116 may be configured to store different types of user activity information in connection with the content items stored in thecontent database114. Examples of user activity information that may be stored in theuser activity database116 may include information indicating scroll activities, scroll time, overall dwell time, dwell time by sections in the given content item, links clicked, position clicked, typing activities, and/or any other activity information indicating various aspects how users viewed and/or interacted with a given content item. For example, the user activity information corresponding to a given content item may indicate user activities engaged by one or more of users during viewing of the given content item. As an illustration, without limitation, the user activity information corresponding to the given content item may include information indicating a first number of specific users that have viewed the given content item, a second number of specific users that have viewed and clicked the given content item, and a third number of specific users that have not viewed given content item. For those users who have viewed the given content item, the user activity information may include information indicating scroll activities, scroll time, overall dwell time, dwell time by sections in the given content item, links clicked, position clicked, typing activities, and/or any other activity information indicating various aspects how the users viewed and/or interacted with the given content item.
• Withsystem10 having been generally described above, various aspects of personalizedcontent recommendation module100 will now be described with reference toFIGS. 2-12.
• FIG. 2 is a flowchart of anexemplary process200 for providing content recommendation, according to various embodiments of the present teaching. As shown, at202, presence of a user may be detected, and a user profile as well as session information associated with the user may be obtained. The user profile obtained at202 may include information indicating content features that have been viewed in the past by the user. For example, such content features in the user profile may include phrases, keywords, specific content identification (e.g., a title of the content item), content topics, content categories and/or any other content features that have been viewed by the user in the past. In some implementations, these features may be indexed in accordance with the frequencies at which the user has viewed them. As an illustration, a given user profile obtained at202 may indicate that the use has viewed contents related to sports (e.g., a topic or category) most often in the past, content items related to finance second most often in the past and so on. As another illustration, the given user profile obtained at202 may indicate the user has viewed content items containing keywords “LeBron James” most often in the past, and contents containing keywords “Cleveland city council” second most often, and so on. The user session information obtained at202 may include information indicating one or more content items that have been viewed by the user in a session or sessions engaged by the user (e.g., the current session and/or a number of historical sessions).
• At204, a first set of candidate content items may be generated based on the user profile obtained at202. The first set of candidate content items may be generated based on content items stored in the content database. The first set of candidate content items at204 may be determined based on considerations, e.g., whether the content item relates to the user's interests according to the user profile. The interests of a user as recorded in the user profile may be determined based on, e.g., declared interests and/or interests that are inferred for the user based on, e.g., user activities observed.
• At208, a second set of candidate content items may be selected from thecontent database114 based on information related to user activities. For example, based on activities of a user on various content items previously presented to the user, the likelihood that the user will be interested in a candidate content item may be estimated. The higher the likelihood estimated, a higher possibility that the user will like to content item. In operation, to select a candidate content item in the second set,system10 may perform the following: for each candidate content item in the content database, determining a set of related content items that are related to the candidate content item and that have been previously presented and viewed by the user; determining a similarity between the candidate content item and each related content item; weighting clicking or non-clicking event by the user during viewing of each related content item with the corresponding similarities; computing the likelihood that the user will click the candidate content item based on aggregation of the weighted clicking or non-clicking events by the user divided by total number of viewing of the related content items by the user; and including the candidate content item in the second set of candidate content items based on the likelihood of the candidate content item (e.g., if it exceeds a predetermined threshold).
• At210, the candidate content items in the first and second sets, as determined at204 and208 respectively, may be ranked. In some implementations, operation(s) at210 may involve ranking the candidate content items in the first and second sets using a learning model.
• At212, the candidate content items may be presented to the user as content recommendations based on their rankings determined at210.
• FIG. 3 illustrates an example of the user-profile based candidate content selection unit106 shown inFIG. 1. It will be described with reference toFIGS. 1-2. As shown in this example, the user-profile based candidate content selection unit106 may be configured to include auser profile processor302, a user-contentfeature vector builder304, a relevance-based content similarity determination module308, a candidate content selection module310, and/or any other module. As shown, theuser profile processor302 may be configured to receive a user profile associated with a user. Such a user profile may include declared interests of the user or inferred interests of the user. The inferred interests of the user may be represented using a feature vector constructed based on, e.g., keywords, phrases, topics, categories, or any other feature associated with content items viewed by the user previously. This feature vector may be constructed based on, e.g., user's activities on the content items previously viewed and acted on. This feature vector may represent inferred interests of the user captured by the values in the vector against different attributes. In some embodiments, the feature vector representing the inferred user interests may be stored in the user profile. In other embodiments, such a feature vector related to inferred user interests may be stored elsewhere and input tosystem10 from a third party vendor (now shown).
• Theuser profile processor302 may be configured to forward the user profile (and the feature vector representing user's inferred interests) to the user-contentfeature vector builder304. The user-contentfeature vector builder304 may then build a user-content feature vector for the given user. The user-content feature vector built by the user-contentfeature vector builder304 may contain components, each of which may indicate that a particular user feature has been viewed by the given user for a corresponding number of times. Such a user-content feature vector is then sent to the relevant-based content similarity determination module308. As illustrated, the relevance-based content similarity determination module308 may be configured to determine the similarity between a content item and the user's interests. For example, when the relevance-based content similarity determination module308 receives the user-content feature vector from the user-contentfeature vector builder304, it may also receive afeature vector306 that provides the average interest of all users with respect to this content item. The average user feature list may be prepared and/or updated based on, e.g., features collected from observations made on all users. This average user feature list may be offline or dynamically trained. As illustrated, the averageuser feature list306 may contain entries indicating features contained in content items (keywords, phrases, topics, etc.) that have been viewed by different users in the past within, e.g., a certain period of time. The average features may be determined based on, e.g., the total number of times being viewed by a user divided by the number of users who have viewed the content item. Based on the averageuser feature list306 and the user-content feature vector, the relevance-based content similarity determination module308 may determine, for each feature, a relative score of the user as compared to the average score from a group of users. The relative score for each feature may be a ratio or any other form that indicates how much the feature from the user deviates from the feature from the group of users representing, e.g., the general population. When the feature from the user is significantly higher than that of the general population, it may be considered that the user is interested in content items that are associated with the feature. A set of features may be selected to guide how to proceed to identify content items that exhibit the features that are ranked high in comparison with the general population.
• Based on the results from the relevance-based content similarity determination module308, the candidate selection module310 selects, from the content database, content items that have the features that are ranked high by the relevance-based content similarity determination module308. For example, the candidate selection module310 may be configured to select content items that contain top features ranked based on the relative scores determined by the relevance-based content similarity determination module308.
• FIG. 4 illustrates anexemplary method400 that may be implemented by the relevance-based content similarity determination module308 shown inFIG. 3. As shown, at402 a user-content feature vector associated with a user may be received. The user-content feature vector received at402 corresponds to a vector for the user constructed based on observation on the behavior of the user. The user-content feature vector may include various attributes or features or components, each of which corresponds to a feature that may occur in content. For example, a feature can be a word, a phrase, or any other features observed from content. Each attribute may have a value which may indicate a number of times a corresponding content including the attribute or feature has been viewed by the user. At404, an average user content feature list may be received. The average user content feature list received at404 corresponds to a user-content feature vector constructed based on observations made on a group of users representing the general population of users. The average user-content feature list may comprise attributes, each of which indicates a number of times content including the corresponding feature has been viewed by the general population (represented by a group of users based on which the average user-content feature list is constructed).
• At406, for each feature in the user-content feature vector, the corresponding number of times the user has viewed content including the feature may be compared with the corresponding average number of times a user from the general population has viewed content with that feature. As an illustration, assuming a user-content feature vector) (f₁. . . f_i. . . f_n) which has attribute or feature values (t₁. . . t_i. . . t_n)_u, respectively, indicating that the user has viewed content including feature f₁in t₁times, viewed content including feature f₂in t₂times, . . . etc. In some examples, the operation(s) at406 may involve normalizing (t₁. . . t_i. . . t_n)_uto (w₁. . . w_i. . . w_n)_u. This may be done by dividing ti with the summation of (t₁. . . t_i. . . t_n). In an average user-content feature list, there are corresponding features and feature values, representing the average number of times the general population has viewed content with each particular feature. The attribute values in the average user-content feature list may also be normalized. Then the operation(s) at406 may divide each w_iin the user-content feature vector by the average value for the corresponding feature, i.e., avg(w_i) in the average user content-feature list to get a sense the strength of the user interests in content including each particular feature as compared with the interests in the same features from an average user.
• At408, top features may be determined for the user based on the comparison at406. The more the features value of the user deviates from that of an average user, more likely it is that the user has a particular interest in content including that feature. At410, a set of candidate content items may be generated based on the top features for the user as determined at408. For example, a set of top features may be determined by selecting those features which have features values that are the most significantly higher than that of the corresponding feature values from the average user-content feature list. Such identified top features may then be used to guide how content is selected from the content database that are deemed to most likely meet the interests of the user. Operation(s) at410 may include querying the content from the content database using the selected top features as key or index to retrieve contents considered to be most relevant to the user interests.
• FIG. 5 illustrates one example of a user-activity based candidatecontent selection unit104 shown inFIG. 1. It will be described in connection withFIG. 1. As shown in this example, the user-activity based candidatecontent selection unit104 may be configured to include a usersession configuration module502, a contentitem information module504, a user contentactivity information module506, an activity based contentsimilarity determination module510, a content item clickprobability determination module508, acandidate selection module512, and/or any other modules. As illustrated herein, the usersession configuration module502 may be configured to receive user session information regarding one or more user sessions associated with a given user. Based on the user session information received, the usersession configuration module502 may determine one or more content items that were viewed by the user during these user sessions. For example, the user session information received may indicate that the user has viewed 4 content items in a user session or that the user has reviewed 7 content items in 3 user sessions. Accordingly, the user sessioncontent determination module502 may identify those content items viewed by the user and forward these content items or identification information thereof to the contentitem information module504.
• The contentitem information module504 may be configured to retrieve such content items and/or information thereof from the content database, such as thecontent database114 shown inFIG. 1.
• The user contentactivity information module506 may be configured to retrieve or receive user activity information with respect to the given content item(s) and analyze the obtained user activity information.FIG. 6 conceptually illustrates one example of user activity information recording activities in association with various content items observed in some users. As illustrated, theuser activity information600 regarding activities of a set of users604a-iduring viewing the content items602a-fmay include information on different types of user activities. As conceptually shown inFIG. 6, information regarding the first illustrated type of user activity may indicate whether users602a-fviewed content items602a-fas illustrated in the first column (from the left) ofFIG. 6. InFIG. 6, when a user did view a content item, there is an edge between the content item and the user (in the first column from the left ofFIG. 6). For example, the edges emanating fromcontent item602aindicates thatusers604a,604b, and604dhave viewed thecontent item602a. Similarly, edges emanating fromitem602bindicate thatusers604b,604d, and604fhave viewed thecontent item602b.
• Another illustrated type of user activity information is related to click-through activity. In the second column ofFIG. 6, it is shown that for each content item, which user have clicked-through the content item. That is, an edge between a content item and a user in the second column ofFIG. 6 indicates that the user clicked-through the content during viewing of the content item. For example, the edge connectingcontent item602aanduser604bindicates that theuser604bclicked thecontent item602aduring viewing of thecontent item602a. Similarly, the edge connectingcontent item602banduser604eand604gindicates that bothuser604eand604ghave clicked thecontent item602bduring viewing of thecontent item602b.
• The user activity information as illustrated in the first and second column from the left ofFIG. 6 indicates whether a user viewing a content item has clicked the content item. Such information can be fused together to reveal integrated information. The fused information can be represented by a fusion graph as illustrated in the third column ofFIG. 6. The fusion graph may be obtained by analyzing information contained in the first and second columns ofFIG. 6. For example, if an edge exists between a content item and a user in both the first and second columns ofFIG. 6, it means that the user not only viewed the content item but also clicked-through the content item, evidencing a stronger interest of the user in the content item. In this case, there will be a solid edge between the user and the content item in the fused graph, as illustrated in the fusion graph inFIG. 6 betweencontent item602aanduser604b. On the other hand, if there is a solid edge between the user and the content item in the first column ofFIG. 6 but without an edge in the second column for the same user and content items, it reveals that although the user did review the content item, the user did not click-through the content item. In this case, in the fused graph, there is a dotted edge between the user and the content item, as illustrated betweencontent item602aanduser604a. It is understood that the user activities that may be captured in theuser activity information600 is not limited to the activities as illustrated inFIG. 6 or engaged by the particular set of user as illustrated. It is also not limited merely to the dimension as illustrated. For example, a session based approach may be employed to capture information related to user activities. In such a session-based approach, activities of all users who are present in a session and whose activities exhibited in the session (e.g., a period of time) may be considered or captured as whole. As an illustration, the total number of clicks on a content item by all users in a session may be considered as the click counts of that content item in the session. Thus, if a number of users viewed a content item but none of them clicked the content item, then there may be a dotted edge between the session and the content item in the third column ofFIG. 6. By comparison, if a number of users viewed a content item and some of them (even only one) clicked then content item, then there may be a solid edge between the session and the content item in the third column ofFIG. 6.
• It should be appreciated that theactivity information600 is not limited to click activities and non-click activities as illustrated in this example.User activity information600 may include other types of user activities. For example, scrolling activities, typing activities, content interaction activities (e.g., whether the user play a game embedded in content item, did the user achieve an objective of the game, etc.) and/or any other user activities during viewing of the content items may be included in theuser activity information600.
• Returning toFIG. 5, the activity based contentsimilarity determination module510 may be configured to determine similarities between content items based on the user activity information obtained and analyzed by the user contentactivity information module506. The operation(s) that may be implemented by the activity based contentsimilarity determination module510 is conceptually illustrated inFIG. 7.
• FIG. 7 conceptually illustrates a similarity between two content items may be determined based on user activity information. As shown in the example,user activity information702aregarding content item #i, for example as obtained and analyzed by the user contentactivity information module506, indicates content item #i was viewed by a first set of users, and whether those users have clicked content item #i during viewing of the content item #i. As also shown,user activity information702bindicates content item #j was viewed by a second set of users, and whether those users have clicked content item #j during viewing of the content item #j. As illustrated, common users in the first set and second set may be obtained. Those common users are the users that have viewed both content item #i and #j. As illustration, suppose those common users may comprise users A, B, C, D, E, F as shown, and their click activities (or lack thereof) during the viewing of both content item #i and #j may be determined. That is, as illustrated, users A and B clicked content item #i when they were viewing it, while users C, D, E, F did not click content item #i when they were viewing it. Similarly, as also shown, users B and D clicked content item #j when they were viewing content item #j, while users A, C, E, F did not click content item #j when they were viewing content item #j. As still illustrated, user activity vectors704a-bmay be generated based on theactivity information702aand702b. As shown, the user activity vector704amay be generated such that each component in the user activity vector704acorrespond to a user A, B, C, D, E, and F, in that order, and the components of the vector704aindicate whether a corresponding user has clicked content item #i. As shown, since users A and B clicked content item #i, and users C, D, E and F did not, the user activity vector704ais generated as (1, 1, 0, 0, 0, 0) with “1” indicating the corresponding user (i.e., users A and B) clicked content item #i and “0” indicating the corresponding users (i.e., users C, D, E, and F) did not click content item #i when they view the content item #i. Similarly, an activity vector704bmay be generated based on theuser activity information702bas illustrated. As will be described below, vectors704a-bmay be used to determine a similarity between items #i and #j.
• FIG. 8 illustrates another example of determining a similarity between two content items #i and #j. As shown in this example, weight factor may be assigned to specific user activities when generating the user activity vector. Different user activities during viewing of the content items #i or #j by the users that viewed both content items #i and #j may be differentiated when generating the user activity vectors. For example, user activities such as click activities, non-clicking activities (i.e., viewed but not clicked content item) scroll activities, dwell time activities (time duration spent on a content item or a portion of the content item) may be accorded with different weights. A number of weighting schemes based on types of user activities and as well as specifics of the user activities may be used. For instance, user click activities may be accorded a first weight W1, user scroll activities may be accorded with weight W2, and user non-clicking activities may be accorded with a negative weight (user viewing of the content item but not clicking may indicate a negative feedback reflecting the user is not interested in the content item). In another example, edges shown inFIG. 6 may be weighted independently based on the specifics of a user activity represented by the edge. For instance, as an illustration, a solid edge may be weighted by the dwell time of corresponding users—i.e., the length of viewing time by the corresponding users on the entire content item or portion(s) of the content item. In another example, a solid edge may be weighted by the scroll activities and links or positions of clicking, during the viewing of the content item by the corresponding users.
• As illustrated, user activity vectors802a-bmay be generated based on the weighted value corresponding to individual ones of the user that have viewed both content items #i and #j. As also illustrated, the user activity vectors802a-bmay be used to determine a similarity between content items #i and #j. Any known methods may be used to compute the similarity between content items #i and #j based on the user activity vectors802a-b. In this example, two of such methods are illustrated. One approach is to compute the similarity based on the cosine between two vectors. Another exemplary approach is to compute a similarity between two vectors by employing a mutual information function.FIG. 9 conceptually illustrates that a similarity between content items may be determined using user activity vectors associated with content items. As illustrated, apoint902 may correspond to an M dimensional user activity vector regarding content item #1 (e.g., theuser activity vector802ashown inFIG. 8). Similarly, apoint904, as illustrated, may correspond to an another M dimensional user activity vector regarding content item #2 (e.g., theuser activity vector802bshown inFIG. 8). The similarity betweencontent items #1 and #2 may be determined by computing a cosine of the angle between vector for content item #1 (corresponding to point902) and vector for content item #2 (corresponding to point904) in the M dimensional space as shown. In implementations, the cosine-based similarity between content items #i and #j may be computed as follows:
• $\mathrm{Sim}\left(i,j\right)=\mathrm{Cos}\mathrm{Sim}\left({\mathrm{vector}}_i,{\mathrm{vector}}_j\right)=\sum _{k=1}^Nw_{i,k}*w_{jk}/\sqrt{\sum _{k=1}^N{w_{i,\mathrm{<figure-callout\; label="k"\; filenames="US20160188725A1-20160630-D00003.png,US20160188725A1-20160630-D00009.png"\; state="\{\{state\}\}">k</figure-callout>}}}^2}\sqrt{\sum _{k=1}^N{w_{j,\mathrm{<figure-callout\; label="k"\; filenames="US20160188725A1-20160630-D00003.png,US20160188725A1-20160630-D00009.png"\; state="\{\{state\}\}">k</figure-callout>}}}^2},,$
• • wherein vector_iand vector_jrepresent user activity vectors associated with the content items #i and #j, respectively, regarding a set of users that viewed both content items #i and #j; and W_i,krepresents a weighted value of a component k in vector_i, the component k corresponding to Kth user in the set of users, and W_j,krepresents a weighted value of a component k in vector_j, W_i,kand W_j,kindicating the Kth user's activity during viewing content items #i and #j, respectively.
• As mentioned, the similarity between content items #i and #j may also be computed based on their respective user activity information using a mutual information function as follows:
• • Sim (i, j)=I (X_iand X_j), wherein X_iand X_jare random variables based on W_i,kand W_j,k,
  • K=1, N (the observations of X_iand X_j), and I denotes a mutual information function that computes the mutual information of variables X_iand X_j
    Other exemplary methods for computing the activity similarity between content items #i and #j may also be employed and are all within the scope of the present teaching. For example, the similarity between content items #i and #j may be computed using a locality sensitive hashing algorithm (LSH) via a hamming distance, a Euclidean distance, or any type of distance between the user activity vectors associated with content items #i and #j.
• FIG. 10 illustrates a flowchart of an exemplary process1000 for determining a similarity between two content items, according to one embodiment of the present teaching. In some implementations, the process1000 may be implemented by an activity based content similarity determination module substantially similar to or the same as activity based content similarity determination module610 as described and illustrated herein. As illustrated, at1002, a first set of users that have viewedcontent item #1 may be determined. At1004, a second set of users that have viewedcontent item #2 may be determined. At1006, user activity information regarding the users in the first set when viewingcontent item #1 may be obtained, and user activity information regarding the user in the second set when viewingcontent item #2 may be obtained. At1008, users that viewed bothcontent item #1 and #2 are determined. At1010, user activity vectors may be generated forcontent items #1 and #2, indicating user activities during the viewing of the content items may be generated. At1012, a similarity betweencontent items #1 and #2 may be generated based on the user activity vectors determined at1008.
• FIG. 11 illustrates the flowchart of another exemplary process1100 for determining a similarity between two content items, according to one embodiment of the present teaching. The difference betweenFIGS. 10 and 11 is that at1110 the user activities of the common users during viewing of thecontent items #1 and #2 may be weighted as described above, and at1112 the user activity vectors forcontent items #1 and #2 may be generated based on the user activities weighted at1110.
• Returning toFIG. 5, the content item clickprobability determination module508 may be configured to estimate the likelihood that a user clicks a given content item.FIG. 12 illustrates a flowchart of an exemplary process for estimating the likelihood that a user will click a given content item. The exemplary process shown inFIG. 12 may be carried out by a content item click probability determination module substantially similar to or the same as the content item clickprobability determination module508 as described and illustrated herein. As illustrated, at1202, a content item #j and as well as information regarding a user may be received. At1204, a set of content items related to content item #j may be determined based on the user information received at1202. This set of related content items may include content items that were viewed by the user during the same session as content item #j, content items related to content item #j based on relevance in accordance with a user profile of the user and were viewed by the user in previous session(s), content items that were specified (e.g., by a provider, administrator, moderator, and/or any other entities related to the system100) as being related to content item #j (e.g., based on a user type of the user and were viewed by the user in previous session(s)), and/or any other content items being considered related to content item #j in accordance with any other criterion. At1206, for each content item in the set determined at1204, a similarity between the content item in the set and content item #j may be determined. The similarities determined at1206 may be expressed as (I_s1, I_s2, . . . I_sN), wherein I_sNdenotes the similarity between Nth related content item in the set and content item #j. At1208, the likelihood that the user will click content item #j may be computed based on the similarities determined at1206. Any known methods may be employed to compute this likelihood. For example, the likelihood that the user will click content item #j may be computed using the following formula:
• $\mathrm{score}=\frac{\sum _{i^{\prime }\in I}s_{i^{\prime }}{\mathrm{act}}_{i^{\prime }}}{\sum _{i^{\prime }\in I}s_{i^{\prime }}}$
• wherein act_idenotes whether the ith related content item in the set was clicked by the user (e.g., the value of act may be either 0 or 1), S_idenotes the similarity between content item #j and the ith related content item, and Score denotes the computed click probability of content item #j by the user. Another example of computing the likelihood of the user clicking content item #j may be computed as follows:
• $\mathrm{score}=\frac{\sum _{i^{\prime }\in I}s_{i^{\prime }}{\mathrm{act}}_{i^{\prime }}+\lambda ·\mathrm{ctr}}{\sum _{i^{\prime }\in I}s_{i^{\prime }}+\lambda }$
• wherein λ denotes a scale factor and ctr denotes the historical click through rate of content item #j. Still in another example, the click probability may be computed as follows:
• $\mathrm{score}=\frac{\sum _{i^{\prime }\in I}s_{i^{\prime }}{\mathrm{act}}_{i^{\prime }}/I+\lambda ·\mathrm{ctr}}{\sum _{i^{\prime }\in I}s_{i^{\prime }}/I+\lambda }$
• wherein |I| denotes the cardinality.
• Referring back toFIG. 1, theunified ranking unit102 as illustrated may be configured to receive a first set of candidate content items as generated by the user-profile based candidate content selection unit106 and a second set of candidate content items as generated by the user-activity based candidatecontent selection unit104. Theunified ranking unit102 may be configured to rank the candidate content items from both sources. The ranking may be carried out in accordance with a learning model derived via themachine learning engine108.
• FIG. 13 is an exemplary diagram of themachine learning engine108 shown inFIG. 1, according to one embodiment of the present teaching. As illustrated herein, themachine learning engine108 may be configured to includestorage1302, a model learning module1308, and/or any other modules. Thestorage1302 may includedata storage1304 for temporally or permanently storing training data received by themachine learning engine108. In some implementations, thetraining data1304 may be received by themachine learning engine108 offline. In some embodiments, thetraining data1304 may include features related to users, features related to content items, user-content cross features, and/or any other features that may be deemed useful in machine learning.FIG. 14 provides some exemplary types of features related to users.FIG. 15 provide a exemplary types of features related to content items.
• The model learning module1308 may be configured to receive the training data, information deemed relevant to a learning model to be trained, and/or parameters to be employed in training the learning model based on the training data. Such training data may reveal much information related to user's interests in content items and is utilized by the present teaching to train a learning model to be used to rank candidate content items. Training data may include user information such as demographics, declared interests, past interests in content items, etc. Training data may also include content information such as classification of content items. Training data may also include event information such as actions users performed on certain content items.
• Features vectors may be generated based on received training data. For example, based on user information as illustrated inFIG. 14, user feature vectors may be generated. Another type of feature vectors may be generated based on information regarding content items, as illustrated inFIG. 15. User information and content information may be cross referenced so that feature vectors may be constructed indicating the user-content cross features. For example, with respect to a user and a content item, whether the user is interested in one or more topics included in the content item may be utilized to construct a user-content cross feature vector, via user activities. If the user clicks a content item having a particular labeled topic, this activity may be converted to a “1” attribute value in a user-content cross feature vector for the feature corresponding to the topic. For each such a user activity, a corresponding user-content cross feature vector may be fed into the model learning module1308 to derive or update a model capturing the preferences of users in content items.
• During the training, the model learning module1308 may use rules configured in the system and stored instorage1306. Such rules and/or the parameters to be used in training a learning model may vary with the learning model adopted. In some embodiments, to train a learning model, the model learning module1308 may generate feature vectors first based on the training data. As discussed above, user information may be used by the model learning module1308 to generate a multiple dimensional feature vector based on various characteristics of the user, e.g., as shown inFIG. 14. For instance, each dimension may be related to a particular user characteristic such as user demographics or user interests. For each dimension, the model learning module1308 may construct a feature vector. For example, given a user's age, the model learning module1308 may construct a user age vector based on a user's manifested age by assigning the user to one of multiple (10) age groups, e.g., age groups corresponding to age ranges of 0-10, 10-20, 20-30, 30-40, 40-50, 50-60, 60-70, 70-80, 80-90, 90-100. As an illustration, assume that an age feature vector is constructed based on the above group of 10 age ranges. If a user's age is 15, the machine learning module1308 may populate the age feature vector for the user as, e.g., (0,1,0,0,0,0,0,0,0,0,0), indicative of the age group that the user belongs to. Based on user information, multiple feature vectors may be constructed. The user interests feature vector may be constructed with respect to category and/or topics of interests based on, e.g., some taxonomy. Such feature vectors map a user into a high dimensional space as a point. Users with similar features will be mapped to points in the high dimensional space with close distances and, hence, indicating that they share similar characteristics. Each such mapped point may be identified based on the identification information of the underlying user.
• Similarly, feature vectors associated with content items may also be constructed for learning purposes. For example, a feature vector related to categories and/or topics of a content item may be constructed. Different topics/categories may be pre-classified into different coded groups. Each content item may have identifiers that indicate the category/topic group(s) that the content item is classified into. To characterize a content item in terms of content category/topic, a feature vector with all possible categories/topics may be constructed with, e.g., each feature value being 1 (belonging to the category/topic) or0 (not belonging to the category/topic). With respect to each content item, its corresponding feature vector is then constructed to have, e.g., “1”s on attributes representing categories/topics that the content item belongs to and “0”s on attributes representing categories/topics that the content item does not belong to. Feature vectors for other characteristics of a content item may also be constructed. That is, for each content item, more than one feature vectors may be constructed and in combination, it maps a content item into a multiple dimensional space as a point. Each point mapped to the content space may be identified based on the identification of the content items. Points mapped close to each other in this content item space indicate that they share certain common or close characteristics.
• Another exemplary type of feature vector that the model learning module1308 may utilize, as discussed above, is user-content cross feature vector. Based on received information on a user and content items associated with the user, the model learning module1308 may construct an integrated feature space that combine the user features with content features. For each user, it may be combined with different content items to generate different feature points in the user-content space. A user-content cross feature vector may also be generated by the model learning module1308 based on each given situation involving a user and a content item. For example, a user-content cross feature vector may be constructed based on different combinations of {user_id, user_feature} and {content_id, content_features} so that the user-content cross feature vector may be constructed based on different ranges of those feature value combinations. In this case, depending on a specific user and a given content item, the model learning module1308 may generate a user-content cross feature vector by assigning a “1” for the user-content feature range appropriate for the given situation and assigning “0”s for all other ranges. For example, if a user is in the age group of 30-40 corresponding to #4 age group and the given content item has a content topic falling in the #13 content category/topic. Such a user-content cross feature vector generated by the model learning module1308 corresponds to a point in the user-content space. Points having a close distance in this user-content space may signify that those users share some common interests in content of certain topics.
• Once the training data points have been mapped to the feature space(s) (user space where users are mapped to via user related feature vectors, content space in which content items are mapped to via content related feature vectors, and user-content space in which user-content cross feature vectors are mapped), the model learning module1308 carries out the learning using such training data based on rules associated with the learning model to be trained, as well the parameters in connection with the learning model to be trained. As all the training data points have been mapped to the feature space, what may be learned include which users share similar interests and on what topics, content items, etc. What may also be learned may include what content items have caused more active involvement of the users who do or do not share similar interests. Such knowledge is learned by the model learning module1308 and embedded in the learning model to be trained. Ultimately, the model learning module1308 may identify how the training data is clustered in the high dimensional space and how the points clustered together may be related to each other and in what manner as well as how data distributed in different clusters may be distinguished based on what features. Such knowledge is captured in the learning process via the learning model and parameters associated therewith so that the learned models can be utilized, in the future, to determine what appropriate content items are to be selected for which user.
• Below, a probabilistic model and learning scheme is disclosed herein. It should be appreciated, however, that any suitable model may be trained by model learning module1308. In addition, any learning model may be adopted to learn the relationship between user information and content information via the linkage established via user activities. One exemplary model for learning is a probabilistic model, as illustrated below:
• • Pr(Click|F)=Φ(βF), where F is feature vectors (e.g., user feature vector, content feature vector, user-content feature vector) translated from user content viewing events and β is the weighting parameter. Φ is the CDF of the standard normal distribution.
    As shown, the probabilistic model may be trained using feature vectors generated based on, e.g., user information, content information, as well as information about user-content interactions. That is, every time a user content viewing event occurs, information about the event may be converted into feature vector(s) as described above, and used by the model learning module1308 to update a current probabilistic model, e.g., with respect to the weights and/or parameters of the underlying model employed. Such modified (updated) model may then be stored and used in the future in subsequent content rankings and/or selections. In this way, a model may be trained in an expectation propagation manner using message passing technique in a factor graph.
• FIG. 16 illustrates an exemplary system diagram showing theunified ranking unit102 shown inFIG. 1. In this illustrated example, theunified ranking unit102 may comprise multiple ranking units—i.e., rankingunits1602a,1602n,1602x, each of which may be configured to independently to rank content items. For example, a ranking unit, e.g.,1602a, may be configured to rank content items based on their popularity in a predetermined period. Theunified ranking unit102 may also include a ranking unit, such as ranking unit1602n, to rank the candidate content items from the first and second candidate content sets, generated based on user profile and user activities, respectively, based on trained learning models. For example, given a trained probabilistic model and the inputs related to a specific user and candidate content items (in both first and second sets), the candidate content items can be ranked using the probabilistic model by computing, for each candidate content item, the probability that the given user will click on the candidate content item. The higher the probability, the high the rank of the candidate content item.
• As illustrated above, one example of the model trained by themachine learning engine108 may include a probabilistic model, which may be used to rank a given candidate content item by estimating a probability of clicking the given content item using Pr(Click|F)=Φ(βF), wherein F denotes features vectors (e.g., user feature vectors, content feature vectors, user-content cross feature vectors, and/or any other feature vectors) generated based on a given content item to be ranked for a user, and Pr(click|F) denotes a probability of the user clicking the given content item as identified by these feature vectors.
• As still shown inFIG. 16, the rankings of candidate content items determined by various ranking units1602 may be sent to thescoring module1604 for generating an integrated ranking score based on the rankings of the candidate content items from various rankingunits1602a, . . . ,1602n, . . .1602x. In some embodiments, thescoring module1604 may employ certain approach to combine the rankings from different ranking units. For example, a weighted sum may be adopted, in which the weights may be applied to the ranking results from different ranking units and such weights may be determined a priori (based on pre-determined preferences of each ranking unit) or learned via, e.g., feedback information. The result from thescoring module1604 may provide an order to the candidate content items sequenced based on their ranking scores. For example, the order can be an ascending order so that the candidate content items with the highest ranking appears first and the candidate content items with lower rankings following that with higher rankings Such an order may be used as the order in which the candidate content items are presented to the user so that the content items that have a high likelihood to be clicked by the user are presented first to improve the possibility that the recommended content meets the interests of the user, determined based on both the user profile as well as the user activities.
• FIG. 17 depicts the architecture of a computing device which can be used to realize a specialized system implementing the present teaching. Such a specialized system incorporating the present teaching has a functional block diagram illustration of a hardware platform which includes user interface elements. The computer may be a general purpose computer or a special purpose computer. Both can be used to implement a specialized system for the present teaching. Thiscomputer1700 may be used to implement any component of the enhanced content recommendation techniques, as described herein. For example, the personalizedcontent recommendation module100, etc., may be implemented on a computer such ascomputer1700, via its hardware, software program, firmware, or a combination thereof. Although only one such computer is shown, for convenience, the computer functions relating to the personalizedcontent recommendation module100 may be implemented in a distributed fashion on a number of similar platforms, to distribute the processing load.
• Thecomputer1700, for example, includes COM ports1750 connected to and from a network connected thereto to facilitate data communications. Thecomputer1700 also includes a central processing unit (CPU)1720, in the form of one or more processors, for executing program instructions. The exemplary computer platform includes aninternal communication bus1710, program storage and data storage of different forms, e.g., disk1770, read only memory (ROM)1730, or random access memory (RAM)1740, for various data files to be processed and/or communicated by the computer, as well as possibly program instructions to be executed by the CPU. Thecomputer1700 also includes an I/O component1760, supporting input/output flows between the computer and other components therein such as user interface elements1780. Thecomputer1700 may also receive programming and data via network communications.
• To implement various modules, units, and their functionalities described in the present disclosure, computer hardware platforms may be used as the hardware platform(s) for one or more of the elements described herein (e.g., personalizedcontent recommendation module210, and/or other components ofsystem100 described herein). The hardware elements, operating systems and programming languages of such computers are conventional in nature, and it is presumed that those skilled in the art are adequately familiar therewith to adapt those technologies to enhance personalized content recommendation described herein. A computer with user interface elements may be used to implement a personal computer (PC) or other input of work station or terminal device, although a computer may also act as a server if appropriately programmed. It is believed that those skilled in the art are familiar with the structure, programming and general operation of such computer equipment and as a result the drawings should be self-explanatory.
• FIG. 18 depicts the architecture of a mobile device which can be used to realize a specialized system implementing the present teaching. In this example, the user device on which content and advertisement are presented and interacted-with is amobile device1800, including, but is not limited to, a smart phone, a tablet, a music player, a handled gaming console, a global positioning system (GPS) receiver, and a wearable computing device (e.g., eyeglasses, wrist watch, etc.), or in any other form factor. Themobile device1800 in this example includes one or more central processing units (CPUs)1840, one or more graphic processing units (GPUs)1830, a display1818, a memory1860, a communication platform1810, such as a wireless communication module,storage1890, and one or more input/output (I/O)devices1850. Any other suitable component, including but not limited to a system bus or a controller (not shown), may also be included in themobile device1800. As shown inFIG. 18, amobile operating system1870, e.g., iOS, Android, Windows Phone, etc., and one ormore applications1880 may be loaded into the memory1860 from thestorage1890 in order to be executed by theCPU1840. Theapplications1880 may include a browser or any other suitable mobile apps for receiving and rendering content streams and advertisements on themobile device1800. User interactions with the content streams may be achieved via the I/O devices1850 and provided to personalizedcontent recommendation module100, and/or other components ofsystem10,
• Hence, aspects of the methods of enhancing ad serving and/or other processes, as outlined above, may be embodied in programming. Program aspects of the technology may be thought of as “products” or “articles of manufacture” typically in the form of executable code and/or associated data that is carried on or embodied in a input of machine readable medium. Tangible non-transitory “storage” input media include any or all of the memory or other storage for the computers, processors or the like, or associated modules thereof, such as various semiconductor memories, tape drives, disk drives and the like, which may provide storage at any time for the software programming.
• All or portions of the software may at times be communicated through a network such as the Internet or various other telecommunication networks. Such communications, for example, may enable loading of the software from one computer or processor into another, for example, from a management server or host computer of personalizedcontent recommendation module100 into the hardware platform(s) of a computing environment or other system implementing a computing environment or similar functionalities in connection with personalizedcontent recommendation module100. Thus, another input of media that may bear the software elements includes optical, electrical and electromagnetic waves, such as used across physical interfaces between local devices, through wired and optical landline networks and over various air-links. The physical elements that carry such waves, such as wired or wireless links, optical links or the like, also may be considered as media bearing the software. As used herein, unless restricted to tangible “storage” media, terms such as computer or machine “readable medium” refer to any medium that participates in providing instructions to a processor for execution.
• Hence, a machine-readable medium may take many forms, including but not limited to, a tangible storage medium, a carrier wave medium or physical transmission medium. Non-volatile storage media include, for example, optical or magnetic disks, such as any of the storage devices in any computer(s) or the like, which may be used to implement the system or any of its components as shown in the drawings. Volatile storage media include dynamic memory, such as a main memory of such a computer platform. Tangible transmission media include coaxial cables; copper wire and fiber optics, including the wires that form a bus within a computer system. Carrier-wave transmission media may take the form of electric or electromagnetic signals, or acoustic or light waves such as those generated during radio frequency (RF) and infrared (IR) data communications. Common forms of computer-readable media therefore include for example: a floppy disk, a flexible disk, hard disk, magnetic tape, any other magnetic medium, a CD-ROM, DVD or DVD-ROM, any other optical medium, punch cards paper tape, any other physical storage medium with patterns of holes, a RAM, a PROM and EPROM, a FLASH-EPROM, any other memory chip or cartridge, a carrier wave transporting data or instructions, cables or links transporting such a carrier wave, or any other medium from which a computer may read programming code and/or data. Many of these forms of computer readable media may be involved in carrying one or more sequences of one or more instructions to a physical processor for execution.
• Those skilled in the art will recognize that the present teachings are amenable to a variety of modifications and/or enhancements. For example, although the implementation of various components described above may be embodied in a hardware device, it may also be implemented as a software only solution—e.g., an installation on an existing server. In addition, the enhanced personalized content recommendation disclosed herein may be implemented as a firmware, firmware/software combination, firmware/hardware combination, or a hardware/firmware/software combination.
• While the foregoing has described what are considered to constitute the present teachings and/or other examples, it is understood that various modifications may be made thereto and that the subject matter disclosed herein may be implemented in various forms and examples, and that the teachings may be applied in numerous applications, only some of which have been described herein. It is intended by the following claims to claim any and all applications, modifications and variations that fall within the true scope of the present teachings.

Claims (20)

We claim:

1. A method, implemented on a machine having at least one processor, storage, and a communication platform connected to a network, for recommending, to a user, content items, the method comprising:

obtaining a user profile characterizing interests of the user;

generating a first set of candidate content items based on the user profile;

generating a second set of candidate content items based on a likelihood that the user clicks a corresponding candidate content item in the second set, wherein each likelihood is estimated based on similarities between the candidate content items in the second set and one or more content items that were previously viewed by the user;

ranking each of the candidate content items in the first set and the second set; and

providing, based on the rankings, the candidate content items in the first and second sets as content recommendations to the user.

2. The method ofclaim 1, wherein the likelihood of a candidate item in the second set is estimated by:

obtaining a set of related content items that are related to the candidate item and that were previously viewed by the user;

determining, for each related content item, a similarity between the related content item and the candidate content item;

determining the likelihood of the candidate content item being clicked by the user by aggregating the similarities between the candidate content item and the related content items; and

including the candidate content item in the second set of candidate content items if the likelihood exceeds a certain threshold.

3. The method ofclaim 2, wherein the step of determining the similarity comprises:

determining a set of common users that viewed both the candidate content item and the related content item;

obtaining information related to one or more activities associated with the set of common users and engaged by the common users during viewing of the candidate content item and the related content item; and

computing the similarity based on the information related to one or more activities associated with the set of common users.

4. The method ofclaim 3, wherein the one or more activities includes at least one of clicking, typing, and/or scrolling.

5. The method ofclaim 3, wherein the step of computing comprises:

generating a first user activity vector for the candidate content item and a second user activity vector for the related content item based on the information; and

estimating the similarity based on the first and second user activity vectors.

6. The method ofclaim 4, wherein the step of estimating is based on at least one of a cosine function, a mutual information algorithm, and a locality sensitive hashing (LSH) algorithm.

7. The method ofclaim 1, wherein the step of ranking comprises ranking each of the candidate content items in the first set and the second set using a learning model, the learning model having been trained with at least one of user feature information, content feature information and user-content cross feature information.

8. A system, implemented on a machine having at least one processor, storage, and a communication platform connected to a network, for recommending, to a user, content items, the system comprising:

a user-profile based candidate content selection unit configured to:

obtain a user profile characterizing interests of the user, and

generate a first set of candidate content items based on the user profile;

a user-activity based candidate content selection unit configured to generate a second set of candidate content items based on a likelihood that the user clicks a corresponding candidate content item in the second set, wherein each likelihood is estimated based on similarities between the candidate content items in the second set and one or more content items that were previously viewed by the user; and

a unified ranking unit configured to

rank each of the candidate content items in the first set and the second set, and

provide, based on the rankings, the candidate content items in the first and second sets as content recommendations to the user.

9. The system ofclaim 8, wherein the user-activity based candidate content selection unit is configured such that the likelihood of a candidate item in the second set is estimated by:

obtaining a set of related content items that are related to the candidate item and that were previously viewed by the user;

determining, for each related content item, a similarity between the related content item and the candidate content item;

determining the likelihood of the candidate content item being clicked by the user by aggregating the similarities between the candidate content item and the related content items; and

including the candidate content item in the second set of candidate content items if the likelihood exceeds a certain threshold.

10. The system ofclaim 8, wherein the user-activity based candidate content selection unit is configured such that determining the similarity comprises:

determining a set of common users that viewed both the candidate content item and the related content item;

obtaining information related to one or more activities associated with the set of common users and engaged by the common users during viewing of the candidate content item and the related content item; and

computing the similarity based on the information related to one or more activities associated with the set of common users.

11. The system ofclaim 10, wherein the one or more activities includes at least one of clicking, typing, and/or scrolling.

12. The system ofclaim 10, wherein the user-activity based candidate content selection unit is further configured such that computing the similarity based on the information related to one or more activities associated with the set of common users comprises:

generating a first user activity vector for the candidate content item and a second user activity vector for the related content item based on the information; and

estimating the similarity based on the first and second user activity vectors.

13. The system ofclaim 12, wherein the similarity is estimated based on at least one of a cosine function, a mutual information algorithm, and a locality sensitive hashing (LSH) algorithm.

14. The system ofclaim 8, wherein the unified ranking unit is configured such that ranking each of the candidate content items in the first set and the second set comprises ranking each of the candidate content items in the first set and the second set using a learning model, the learning model having been trained with at least one of user feature information, content feature information and user-content cross feature information.

15. A machine-readable tangible and non-transitory medium having information for recommending, to a user, content items, wherein the information, when read by the machine, causes the machine to perform the following:

obtaining a user profile characterizing interests of the user;

generating a first set of candidate content items based on the user profile;

generating a second set of candidate content items based on a likelihood that the user clicks a corresponding candidate content item in the second set, wherein each likelihood is estimated based on similarities between the candidate content items in the second set and one or more content items that were previously viewed by the user;

ranking each of the candidate content items in the first set and the second set; and

providing, based on the rankings, the candidate content items in the first and second sets as content recommendations to the user.

16. The machine-readable and non-transitory medium ofclaim 15, wherein the likelihood of a candidate item in the second set is estimated by:

obtaining a set of related content items that are related to the candidate item and that were previously viewed by the user;

determining, for each related content item, a similarity between the related content item and the candidate content item;

determining the likelihood of the candidate content item being clicked by the user by aggregating the similarities between the candidate content item and the related content items; and

including the candidate content item in the second set of candidate content items if the likelihood exceeds a certain threshold.

17. The machine-readable and non-transitory medium ofclaim 15, wherein the step of determining the similarity comprises:

determining a set of common users that viewed both the candidate content item and the related content item;

obtaining information related to one or more activities associated with the set of common users and engaged by the common users during viewing of the candidate content item and the related content item; and

computing the similarity based on the information related to one or more activities associated with the set of common users.

18. The machine-readable and non-transitory medium ofclaim 17, wherein the one or more activities includes at least one of clicking, typing, and/or scrolling.

19. The machine-readable and non-transitory medium ofclaim 17, wherein the step of computing comprises:

generating a first user activity vector for the candidate content item and a second user activity vector for the related content item based on the information; and

estimating the similarity based on the first and second user activity vectors.

20. The machine-readable and non-transitory medium ofclaim 15, wherein the step of ranking comprises ranking each of the candidate content items in the first set and the second set using a learning model, the learning model having been trained with at least one of user feature information, content feature information and user-content cross feature information.

Images